AI Agent News
Track major events, funding, model releases and breakthroughs across the AI Agent landscape
Latest industry news
Track major events, funding, model releases and breakthroughs across the AI Agent landscape
Key events timeline
OpenClaw erupts on GitHub
OpenClaw hits global GitHub Top 10 in 10 days, outpacing the Linux kernel star growth
Meta acquires Manus for $2B
Meta acquires Manus AI for $2B, locking in the general-purpose Agent race
DeepSeek-V3 open-sourced
The value king, at just 5% of GPT-4 cost
Manus goes viral overnight
The world's first general-purpose AI Agent draws unprecedented attention
OpenAI Deep Research
OpenAI ships a deep-research Agent that generates professional reports in one click
MCP Servers pass 500
The MCP ecosystem erupts — 500+ servers built in 3 months
DeepSeek-R1 stuns the world
Open-source reasoning model at just 3% of OpenAI cost, reshaping the global AI landscape
MCP protocol born
Anthropic releases the Model Context Protocol, the de facto standard for Agent interfaces
Claude Computer Use
Anthropic lets AI directly control the computer screen for the first time, opening a new paradigm
Replit Agent full-stack automation
Natural language to a shipped product, aimed at non-engineers
Cursor ARR passes $100M
The fastest-growing SaaS ever, the new king of AI coding tools
Claude 3.5 tops SWE-bench
The strongest coding AI, bug-fixing at a junior engineer level
Devin launches
The world's first autonomous AI software engineer, able to complete full coding tasks on its own
Infinigence Unveils Agentic Infra Strategy: Inference Cost Down 10x in One Year, Daily Token Calls Up 40x
At WAIC 2025, Infinigence officially launched its "Front Shop, Back Factory, One Center" Agentic Infra strategic layout, aiming to build autonomous AI infrastructure for the Agent era. The strategy comprises a Computing Power Distribution Center (Agentic Infra platform), a Token Factory (Agentic MaaS), and an AI Productivity Store (industry solutions), with the core formula: AI Productivity = Intelligent Resource Scale × Token Conversion Efficiency × AI Productivity Conversion Efficiency. ## Key Technologies & Data - **Cross-Cluster Heterogeneous PD Separation Architecture (PDD)**: Splits Prefill and Decode onto different chips, introduces a RelayDecode middle layer, reducing first-token latency by 51.5% and per-token cost by 37.5%. - **Inference Cost Down 10x in One Year**: Through full-stack optimization, unit token inference cost dropped 10x in one year, with another 10x reduction expected. - **Daily Token Calls Up 40x**: As of July 2025, Agentic MaaS platform daily token calls increased 40x compared to December 2024. - **Computing Power Distribution Center**: Deployed and reached over 37,000P computing power, covering 16 mainstream chips, completed China's first cross-region hybrid training over 4,000 km, with a 165% performance improvement. - **Cross-Cluster Reinforcement Learning**: Achieved stable operation for one week without interruption, future support for clusters of 100K+ cards. ## Infrastructure Agent Swarm - **Platform Steward Agent**: Independently solves 80% of daily issues, improving O&M efficiency by 5x. - **Intelligent Computing Cluster O&M Agent**: Key fault handling efficiency improved by 6x. - **High-Performance Operator Generation Agent KernelMind**: In GLM 5.2 model tests, NVIDIA flagship card performance improved by 7.3%, AMD flagship card by 39%. ## Industry Collaboration & Ecosystem Infinigence has deeply collaborated with leading LLM companies such as Kimi, Zhipu AI, MiniMax, and StepFun, jointly launched the "Tianwen" model service portal with Shanghai Mobile, and the TokenDance platform for developers with Guancha. Industry solutions cover entertainment/gaming, healthcare, legal, energy, and more. Infinigence CEO Xia Lixue stated that the core of Agentic Infra is to make infrastructure itself an agent that can work, supporting the online evolution of next-generation Agentic AI through three barriers: full-chain, AI-native, and full coverage.
Peking University Team Develops Millisecond Neuromorphic Dynamics Chip Based on Phase-Change Memristors, Accelerating Cortex Reconstruction by Nearly 480x
On July 2, 2025, a team led by Yang Yuchao from Peking University, in collaboration with Song Zhitang's group from the Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, published a paper in *Science* reporting the world's first millisecond neuromorphic dynamics system chip based on phase-change memory (PCM) memristors. Fabricated in 40nm technology, the chip leverages the controllable conductance drift of PCM devices for in-situ step-size search and integrates compute-in-memory to accelerate neural network inference, reducing the single iteration latency of neuromorphic dynamics to 2.12 milliseconds. In brain cortex surface reconstruction tasks, it achieves 50.38x to 478.18x speedup and 11.75x to 24.73x power reduction compared to an NVIDIA A100 GPU. ## Background and Challenges Neuromorphic dynamics systems (NDS) embed neural networks into differential equation solvers, offering higher accuracy than traditional methods in high-fidelity geometric modeling (e.g., cortex reconstruction), but suffer from slow computation. Traditional digital hardware (GPUs, ASICs) is limited by the memory wall and frequent data movement, with single iterations taking hundreds of milliseconds. Key bottlenecks include high-cost adaptive step-size search, large latency in embedded neural network (ENN) inference, and lack of dedicated support for special computation kernels (e.g., square root). ## Core Innovations - **Controllable Conductance Drift (CCD)**: The team discovered that the conductance of PCM devices drifts predictably and controllably over time, and used this to encode the step size Δt, turning step-size search into a physical evolution process, eliminating read/write and multiplication operations in traditional digital circuits. - **Compute-in-Memory (CIM)**: Utilizing the multi-level conductance of PCM, ENN weights are programmed into memristor arrays, and multiply-accumulate operations are performed in the analog domain in situ, avoiding data movement. A dual-column differential structure supports 16 conductance states (8 levels), covering weight matrices from 32×32 to 128×128. ## Chip Architecture and Performance The chip is fabricated in 40nm technology, runs at 50 MHz, and contains a 288×512 PCM 1T1R array (approximately 147,000 devices). The combined area of compute-in-memory and step-size drift arrays is only 0.28 mm². A single NDS iteration (fourth-order Runge-Kutta) consists of four steps: ENN compute-in-memory, conductance drift step-size search, intermediate quantity combination, and output with error estimation. Compared to a dedicated ASIC, it achieves 3.82x to 36.27x speedup and 11.75x to 24.73x power reduction. ## Application Validation: Cortex Reconstruction The team applied the chip to reconstruct white matter and gray matter cortical surfaces, generating non-self-intersecting closed manifold meshes. It meets high-fidelity requirements in terms of average symmetric surface distance (ASSD) and Hausdorff distance (HD). Compared to FreeSurfer (8722–11860 seconds) and GPU A100 (1.83–21.47 seconds), the PCM-NDS chip reduces processing time to 3.85 ms to 426.31 ms, achieving a maximum speedup of 478.18x. ## Impact and Evaluation This work pushes real-time computation of neuromorphic dynamics systems to the millisecond level for the first time, providing a hardware foundation for brain-computer interfaces, brain digital twins, and neuronavigation. *Science* published a concurrent commentary, calling it "a paradigm shift toward physics-driven computing."
Lilian Weng Proposes Harness Engineering for AI Self-Improvement: Near-Term Path Lies in Peripheral Systems, Not Model Weights
Former OpenAI VP of Safety Research and co-founder of Thinking Machines Lab, Lilian Weng, published a blog post on July 4, 2026, titled "Harness Engineering for Self-Improvement," systematically outlining a near-term feasible path for recursive self-improvement (RSI) in AI. She argues that AI self-improvement should not start with models directly rewriting their own weights, but rather prioritize optimizing the Harness system that wraps around the model—the execution layer responsible for orchestrating model thinking, tool invocation, context management, and result evaluation. This view was endorsed by DeepSeek researcher Tianyi Cui. ## Core Argument: Harness Layer Over Model Weights Weng points out that the near-term path for RSI is more likely to occur at the Harness layer rather than the model weight layer. The Harness is a system built around the base model, determining how the model plans, invokes tools, manages context, stores outputs, and evaluates results. She believes that improving the "mechanism for obtaining better answers" is more realistic than improving the answers themselves. Many improvements at the Harness layer may eventually be internalized as model capabilities, but the external interface should be retained. ## Three Design Patterns for Harness Weng identifies three core design patterns for Harness engineering: - **Workflow Automation**: Define goal-oriented loops (plan → execute → observe/test → improve), where the model iterates by analyzing its own trajectory at runtime, rather than relying on static prompts. - **File System as Persistent Memory**: Store outputs such as experiment logs and code diffs as files to avoid context window overflow, leveraging the LLM's bash read/write capabilities. - **Sub-Agents and Backend Tasks**: Spawn multiple sub-agents for parallel execution, monitored and merged by a process manager to ensure state recoverability. ## Harness Optimization Path: From Context to Evolutionary Search Weng summarizes the evolution of optimization targets as: instruction prompts → structured context → workflows → Harness code → optimizer code. Specifics include: - **Context Engineering**: ACE treats context as a dynamic manual; MCE separates context management skills from content. - **Workflow Design**: AI Scientist, ADAS, AFlow, and others incorporate process structure into the search space. - **Self-Improving Harness**: Self-Harness automatically optimizes Harness configurations through a cycle of weakness mining, proposal generation, and validation. - **Evolutionary Search**: Darwin Gödel Machine (DGM) allows coding agents to directly modify their own Harness code, improving from 20% to 50% on SWE-bench Verified and from 14.2% to 30.7% on Polyglot, matching or exceeding human-designed systems. ## Challenges and Boundaries Weng candidly lists current bottlenecks: - Evaluators are too weak; only tasks like code and math that can be automatically evaluated run the cycle smoothly, while scientific taste and other qualities are hard to quantify. - Lifecycle management of context and memory remains unsolved. - Negative results are systematically ignored; models may not be good at judging when to abandon hypotheses. - Diversity collapse: evolutionary loops may converge to local optima. - Reward hacking remains an issue. ## Connection to Scaling Laws Two weeks ago, Weng published "Scaling Laws, Carefully," pointing out that fitting scaling laws is affected by details such as parameter counting methods, loss precision, and data duplication, making extrapolation risky. Both articles together suggest: when the marginal benefit of stacking parameters and data diminishes, Harness engineering may become the second growth curve for AI.
Lilian Weng Proposes New Path for AI Self-Improvement: Harness Engineering May Precede Model Weight Optimization
Lilian Weng, former VP of Safety Research at OpenAI and co-founder of Thinking Machines Lab, published a blog post titled "Harness Engineering for Self-Improvement," systematically outlining a practical path for recursive self-improvement (RSI) in AI. She argues that in the near term, AI self-improvement may not start with directly modifying model weights, but rather with the Harness layer—the execution system that orchestrates model reasoning, tool calls, context management, and result evaluation. ## Core Thesis: Harness as a Viable Starting Point for RSI Weng defines Harness as "the execution system built around the base model," which determines how the model observes the environment, invokes tools, manages context, stores state, and evaluates results. She points out that the success of coding agent products like Claude Code and Codex has already demonstrated the critical role of Harness in AI deployment. Compared to early agent frameworks (LLM + memory + tools + planning + action), Harness engineering adds workflow design, evaluation, permission control, and persistent state management, making it closer to runtime and software system design. ## Optimization Path: From Context to Harness Code Weng outlines a progressive chain of Harness optimization: prompt → structured context → workflow → Harness code → optimizer code. Specific approaches include: - **Context Engineering**: ACE (Agentic Context Engineering) treats context as a continuously updated "operating manual," maintained by three components: generator, reflector, and curator. MCE (Meta Context Engineering) further splits "how to manage context" and "context content" into two layers of optimization. - **Workflow Design**: Systems like AI Scientist, ADAS, and AFlow treat agent workflows themselves as searchable optimization objects, evolving from human-designed processes to model-involved design, and finally to workflow structures entering the search space. - **Self-Improving Harness**: The Self-Harness system enables the model to analyze its own failure modes and modify Harness configurations through three steps: weakness mining, Harness proposal, and proposal validation. Tests on models like MiniMax M2.5, Qwen3.5, and GLM-5 show that this method can learn differentiated configurations tailored to each model's weaknesses. - **Evolutionary Search**: DGM (Darwin Gödel Machine) allows coding agents to directly modify their own Harness code repository. Experiments show that with Claude 3.5 Sonnet as the base, starting from a simple initial configuration, the evolved DGM agent improved from 20% to 50% on SWE-bench Verified and from 14.2% to 30.7% on Polyglot, reaching or exceeding human-designed levels. ## Challenges and Boundaries Weng candidly lists current bottlenecks for RSI: - Evaluators are too weak and vague; only tasks with clear objective feedback (e.g., code, math) can currently run self-improvement loops. - Lifecycle management of context and memory: the more autonomous the task, the more memory needs to be managed. - Negative results are systematically ignored; models trained on success-dominated data may not be good at judging when to abandon hypotheses. - Diversity collapse: evolutionary and reinforcement learning loops tend to converge to local optima. - Allowing programs to modify system-level code breaks abstraction boundaries, and reward hacking remains a problem. ## Long-Term Outlook: Harness Improvements May Be Internalized Weng believes that Harness engineering will evolve toward a "meta-methodology"—optimizing the "mechanism for obtaining better answers" itself. A mature Harness can support the research loop of model self-improvement, while smarter models can prevent over-engineering of the Harness. Eventually, many Harness-level improvements may be internalized into core model behavior, but external context and tool interfaces are likely to remain. This pattern has precedent in prompt engineering history: as instruction tuning and reasoning capabilities improve, manual prompt tricks become less central, but the need to specify goals, constraints, context, and evaluation does not disappear.
Lilian Weng's Two Recent Blog Posts: Harness Engineering May Become a New Paradigm for AI Self-Improvement, Foundations of Scaling Law Shaken
Lilian Weng, former VP of Safety Research at OpenAI and co-founder of Thinking Machines Lab, published two technical blog posts within two weeks, discussing the limitations of Scaling Laws and the potential of Harness Engineering for AI self-improvement. The first post, "Scaling Laws, Carefully," points out that the fitting results of classical Scaling Laws are highly sensitive to details such as parameter counting methods and loss function precision. The disagreement between Kaplan and Chinchilla stems from bookkeeping-level differences, and data repetition leads to overfitting, undermining the premise of "infinite data." The second post, "Harness Engineering for Self-Improvement," systematically reviews design patterns for the execution system (Harness) wrapped around the base model, including workflow automation, file system persistent memory, sub-agent parallelism, etc. It argues that recursive self-improvement (RSI) is more likely to occur at the Harness layer in the near term rather than through direct rewriting of model weights. The post also introduces context optimization methods like ACE and Meta Context Engineering, as well as self-optimization frameworks such as Self-Harness and Darwin Gödel Machine, while candidly listing unresolved challenges including evaluation ambiguity, memory management, diversity collapse, and reward hacking. Together, the two posts point to a trend: when the marginal benefit of scaling model parameters diminishes, Harness Engineering may become a second growth curve for AI capability enhancement.
DeepSeek DSpark Technology Ported to Apple Silicon, Boosts Local Mac LLM Speed by Up to 60%
DeepSeek's speculative decoding technology DSpark, open-sourced on June 27, has been ported to Apple Silicon (Mac) by engineer Abdur Rahim under the project name mlx-dspark. The port supports Gemma-4 12B and Qwen3-4B models, achieving speedups of approximately 1.6× (from 18.4 tok/s to ~30 tok/s) and 1.4× (from 52.9 tok/s to ~73 tok/s) on M4 Pro, respectively. ## Technical Principles and Implementation - DSpark core idea: A small draft model quickly generates candidate tokens, which are then batch-verified by the target model. If accepted, they are kept; otherwise, regeneration occurs. - Cost differences on Apple Silicon: Data center GPUs have fixed batch verification costs, while Apple Silicon's verification cost scales linearly with the number of candidate tokens. Rahim measured that each additional token verification for Gemma-4 12B takes about 14 ms, and built a cost model yielding a theoretical speedup upper bound of ~2.2×. - Implementation details: The draft model is extracted from HuggingFace checkpoints and quantized to 4-bit (only 1.8GB). The target model defaults to 8-bit precision (bf16 verification is more expensive and actually slower). ## High-Fidelity Reproduction and Sampling Support - Most local ports only support greedy decoding, but mlx-dspark implements the temperature sampling method from the paper, producing output distributions strictly identical to the target model, byte-for-byte. - Rahim found: When the draft model is paired with a non-instruction-tuned base target model, the candidate acceptance rate is only 47%; switching to an instruction-tuned version raises the acceptance rate to 82%. ## Integration of DFlash Scheme - At the request of DFlash paper author Jian Chen, Rahim integrated the DFlash scheme into mlx-dspark. DFlash uses parallel block diffusion (generating 16 tokens at once), achieving a speedup of ~2.1× (~36 tok/s) on code and math tasks, outperforming DSpark. - However, in open-ended chat scenarios, DFlash's acceptance length is limited, making DSpark faster. mlx-dspark v0.0.3 allows users to manually adjust DFlash's effective block length to suit different tasks. ## Impact and Outlook - This is the first native Apple Silicon implementation of DSpark since its open-source release, enabling Mac users to enjoy acceleration without relying on data center GPUs. - Rahim indicates the method can scale to larger draft models (e.g., Qwen3-8B and 14B). - Concurrently, DeepSeek is actively recruiting, including Tsinghua University PhD student Gu Yuxian (Apple PhD Scholar), whose research covers model compression and efficient architectures, synergizing with technologies like DSpark.
MiningLamp Open Sources Agent Internet Platform Octo, Building a New Paradigm of Human-Machine Collaboration
MiningLamp recently open-sourced the Agent Internet platform Octo, aiming to address the pain point of current AI Agents operating in silos and lacking collaboration. The platform aggregates scattered Bots into a unified collaborative space, supporting efficient connection and communication between humans and Agents, as well as among Agents themselves. It introduces Matter and Taste mechanisms for tracking complex tasks and accumulating organizational preferences. Octo offers six collaboration modes (Solo, Roundtable, Critic, Pipeline, Split, Swarm), covering scenarios from single-person tasks to multi-Agent pipelines and competitive selection. The product includes Web App, mobile app, browser extension, and CLI, bringing Agents to where work happens. The project is open-sourced on GitHub.
StepFun and Multiple Universities Open-Source JetSpec, Achieving Up to 9.64x Speedup in Speculative Decoding
StepFun, in collaboration with teams from UC San Diego, Zhejiang University, University of Illinois, and Nanjing University, has recently open-sourced JetSpec, a speculative decoding framework. JetSpec introduces a causal parallel draft head that generates an entire candidate tree in a single forward pass while maintaining branch-level causal consistency, breaking through the scaling ceiling of traditional speculative decoding. ## Core Speedup Results On H100 GPUs, JetSpec achieves significant acceleration on the Qwen3-8B model: - **MATH-500**: End-to-end decoding speedup of **9.64x**, with an average accepted length of 10.76 tokens - **GSM8K**: 7.82x - **AIME25**: 8.78x - **HumanEval**: 7.12x - **MBPP**: 6.73x - **LiveCodeBench**: 7.67x - **MT-Bench**: 4.58x On the MoE model Qwen3-30B-A3B, MATH-500 achieves a 9.45x speedup and AIME25 reaches 9.35x, demonstrating that the method is architecture-agnostic. ## Technical Principle: Resolving the Causality-Efficiency Dilemma The speedup of speculative decoding is limited by the draft generation cost (c) and the per-token acceptance rate (α). Traditional methods face a dilemma: - **Autoregressive drafts (e.g., EAGLE series)**: Good causal consistency and high acceptance rates, but serial generation steps increase draft cost with tree depth. - **Block parallel drafts (e.g., DFlash series)**: Generate multiple candidates in one forward pass with very low cost, but lack branch-level causal conditioning, leading to "locally plausible, globally inconsistent" outputs and low acceptance rates. JetSpec's **causal parallel draft head** reuses frozen hidden states from the target model and employs a **tree-causal attention mask**: each tree node can only see the original prefix and its branch's ancestor tokens. All nodes are computed in parallel in a single forward pass while maintaining autoregressive dependencies within branches. Training uses forward KL divergence distillation on a dataset comprising 780K samples from the Nemotron Post-Training Dataset V2 and 20K samples from CodeAlpaca. ## Complementarity with Contemporary Work DSpark JetSpec and DeepSeek's contemporaneous open-source DSpark optimize inference efficiency from different angles: - **DSpark**: Targets high-concurrency, budget-constrained scenarios, improving throughput via lightweight correction heads and confidence scheduling. - **JetSpec**: Targets low-latency, compute-budget-rich scenarios, maximizing single-verification acceptance rate through causal parallel tree generation to reduce single-user latency. Both point to causality as the key to next-generation speculative decoding. ## Service Scenarios and Budget Strategies - **Low concurrency (batch size=1)**: Increasing tree budget from 16 to 128 raises throughput from 443.3 TPS to 968.2 TPS, with speedup increasing from 3.09x to 6.75x. - **High concurrency (batch size=32)**: At budget 256, speedup drops to 2.85x; small to medium budgets are recommended. The team currently evaluates only static budgets; dynamic adjustment is left for future work. ## Open Source and Team The project is open-sourced on GitHub ([hao-ai-lab/JetSpec](https://github.com/hao-ai-lab/JetSpec)), the paper is on arXiv ([2606.18394](https://arxiv.org/abs/2606.18394)), and model weights are on Hugging Face ([JetSpec](https://huggingface.co/JetSpec)). Authors include StepFun CEO Jiang Daxin, CTO Zhu Yibo, and UCSD PhD student Lanxiang Hu, among others.
Renmin University and Microsoft Open Source Autonomous Research Framework Arbor: Structured Search Based on Hypothesis Trees Achieves State-of-the-Art on All Six Tasks
The Gaoling School of Artificial Intelligence at Renmin University of China, in collaboration with Microsoft Research, has open-sourced the autonomous research framework Arbor. It addresses the challenges AI agents face in long-term research tasks, such as difficulty accumulating experience and blind trial-and-error. Arbor's core is the Hypothesis Tree Refinement (HTR) mechanism, which organizes the research process as a continuously growing tree where each node contains a hypothesis, code version, experimental evidence, and distilled insights. The system adopts a Coordinator-Executor two-tier architecture: the Coordinator handles global strategy, maintains the hypothesis tree, and decides exploration directions; the Executor runs specific experiments in isolated environments and returns structured reports. On six real research tasks (covering model training, Harness engineering, and data synthesis), Arbor achieved the best results on held-out test sets, with an average held-out gain over 2.5 times that of Codex and Claude Code. On MLE-Bench Lite with GPT-5.5, Arbor's Any Medal score reached 86.36%, the highest to date. Ablation studies show that removing the tree structure or disabling insight feedback leads to significant performance drops, verifying the necessity of combining tree and insights. Arbor's token consumption is comparable to baseline methods (approximately 20M–43M tokens), indicating that gains come from structured search rather than increased computation. The project is open-sourced, including a standalone CLI and Agent Skill, supporting use in environments like Codex/Claude Code.